OpenAI launches text watermarking under the EU AI Act

OpenAI's textGrain system aims to meet Article 50 of the EU AI Act, which requires generative AI providers to mark synthetic outputs. This phased rollout introduces statistical provenance tracking for machine-generated text, initially focusing on the European Union while offering an optional feature to global enterprise API clients.

The effectiveness of statistical watermarks, however, faces significant challenges outside controlled environments. OpenAI's internal tests indicate that detection accuracy drops considerably with shorter text samples, structured content like mathematics, or even minor alterations. Replacing just 10% of words with synonyms can reduce detection rates from 92% to 66%, and swapping 25% reduces it to 17%.

The inherent inaccuracies in AI detection tools, particularly statistical watermarking, pose an operational concern. While the EU AI Act mandates machine-readable provenance, the probabilistic nature of natural language detection creates a tension between legal requirements and technical realities. Relying on closed, commercial detection tools also raises questions about public transparency standards, as their internal evaluation processes remain unaudited.